Structure of Arrays vs Array of Structures in CUDA

Question

From some comments that I have read in here, for some reason it is preferable to have Structure of Arrays (SoA) over Array of Structures

Answer 1

SoA is effectly good for SIMD processing. For several reason, but basically it's more efficient to load 4 consecutive floats in a register. With something like:

 float v [4] = {0};
 __m128 reg = _mm_load_ps( v );

than using:

 struct vec { float x; float, y; ....} ;
 vec v = {0, 0, 0, 0};

and create an __m128 data by accessing all member:

 __m128 reg = _mm_set_ps(v.x, ....);

if your arrays are 16-byte aligned data load/store are faster and some op can be perform directly in memory.